To investigate Natural Language Processing combined with AI enabled image analysis tools applied to PETCT reporting to enhance specificity in follow-u

本研究的目的是使用自然语言处理（NLP）从PET-CT报告中提取癌症信息。PET-CT报告使用复杂的受试者特定语言，可能需要其他放射科医生、肿瘤科临床医生和患者准备好并理解。这些小组的专业知识水平各不相同，确定一份报告的总体判断可能并不明确。这项研究旨在利用过去的报告来构建一个AI分类器，该分类器可以根据TNM癌症分期标准标记未来的报告。训练数据在报告级别进行注释，以更好地考虑治疗背景。肿瘤、淋巴结和转移结果有单独的文档级分类。附带的发现也被注释，看看这些是否可以在跨度级别提取，因为这些临床细节可能会由于报告的自由文本格式而丢失。成功提取这些标签和发现将允许更容易地构建其他项目的数据集，以及其他医疗保健分析。特别是：“正常”扫描（没有癌症异常）的可靠提取可以通过确定反映这种分类的图像特征来帮助进一步研究。我们还将探索NLP的最新突破如何改善放射学报告摘要（或自动印象生成），因为PET-CT尚未彻底探索这项任务。该方法是实验性的;根据公认的评估指标测试模型的设计和应用。每个类别的精确度，召回率和F1分数对于证明其有效性至关重要。受试者工作曲线下面积也应该用于评估，因为它使用的是类别的概率，而不是单独的最终分类。Rouge a文本总结指标也将用于评估偶然发现和印象生成。我还将探索潜在的人类评估技术。

The goal of this study is to use Natural Language Processing (NLP) to extract information from PET-CT reports for cancer. PET-CT reports use complex, subject-specific language which potentially needs to be ready and understood by other Radiologists, Oncology Clinicians and Patients. These groups have different levels of expertise and ascertaining the overall judgement of a report can be unclear. This study seeks to use past reports to build an AI classifier which can label future reports based on TNM cancer staging criteria. The training data is annotated at report level to better take into account the treatment context. There are separate document-level classifications for Tumour, Node and Metastasis findings. Incidental findings are also being annotated to see if these can be extracted at span-level as these clinical details can be missed due to the free text format of reports.Success in extracting these labels and findings would allow for easier construction of datasets for other projects, and other healthcare analysis. In particular: The reliable extraction of 'normal' scans (with no cancerous abnormalities) could help further research by ascertaining image features that reflect this classification. We will also seek to explore how recent breakthroughs in NLP can improve radiology report summarisation (or automatic impression generation) as this task has not been explored thoroughly with regard to PET-CT.The methodology is experimental; testing the design and application of the models against accepted evaluation metrics. Precision, Recall and F1 score for each class would be essential to demonstrate its effectiveness. Area under the receiver operating curve should also be used for evaluation as it uses the probabilities of the classes, as opposed to the final classification alone. Rouge a text summarisation metric will also be used to evaluate incidental findings and impression generations. I will also explore potential human evaluation techniques.